Source | SourceSight

Post-Spring Festival, the popularity of DeepSeek remains unabated. To expedite the industry-wide adoption of large models, the DeepSeek series has been deployed across multiple leading cloud platforms globally.

Simultaneously, DeepSeek has shattered long-held beliefs in the large model sector, ushering in an AI revolution centered on "low cost and high efficiency".

A stark example is that, compared to traditional AI training models, DeepSeek significantly reduces the demand for computing resources under similar tasks, thereby lowering hardware investment costs. This is the confidence behind DeepSeek's ability to offer services comparable to ChatGPT-o1 without user fees.

As a result, 2025 has emerged as a year of continuous cost reduction and efficiency enhancement for AI large models. Inspired by DeepSeek, numerous domestic and international vendors have announced the launch of DeepSeek large models, offering genuine value through cost-effective services to attract more users.

As a domestic AI pioneer, Baidu stands out as particularly noteworthy.

01

Moore's Law in Large Model Effectiveness

To find a platform deeply involved in the commercialization of large models in China, Baidu is undoubtedly one of the key players.

On February 3, Baidu Intelligent Cloud announced that its Qianfan platform officially listed the DeepSeek-R1 and DeepSeek-V3 models, introducing an ultra-low-price plan – prices are only 30% of the official list price for DeepSeek-V3 and 50% for DeepSeek-R1. Additionally, users can enjoy limited-time free services.

Baidu Intelligent Cloud stated that the models accessed this time have fully integrated the Qianfan inference chain, incorporating Baidu's exclusive content security operators to enhance model security and provide enterprise-level high availability guarantees. It also supports comprehensive BLS log analysis and BCM alerting, helping users build intelligent applications safely and stably.

According to public data, the overall cost of DeepSeek-R1 is approximately 1/30 of that of the OpenAI GPT-1 model. This data has caused a stir, shaking the beliefs of practitioners worldwide.

Professor Matsuo Toyohiko, a Tokyo University professor renowned as the "pioneer of Japanese AI research," recently stated that DeepSeek technology is exceptional, with performance close to that of American OpenAI, which operates "ChatGPT." Each new AI model release is accompanied by a detailed paper outlining the technology used and the meticulous improvements implemented.

Matsuo also praised DeepSeek for its open-source approach. Its latest model "R1," released in January, underwent reinforcement learning to enhance reasoning ability, ultimately demonstrating reasoning performance equivalent to OpenAI's GPT-1.

Concurrently, providing a more convenient, efficient, and economical product experience has become crucial for AI large model enterprises to excel in competition.

Behind Baidu Intelligent Cloud's Qianfan platform's ultra-low-price plan lies its effort to reduce users' AI model trial-and-error costs, aligning with the current trend of popularizing AI technology.

According to data released by market research and consulting firm Gartner, by 2027, over half of the AI models used by enterprises will have specific industry or business functions, compared to less than 1% in 2023.

However, from model training to application development, the high investment costs and unobvious short-term benefits caused by consuming large amounts of funds and stacking computing power are the main challenges for enterprises to realize business scenario-based implementation of large models. Reducing technology costs is the primary driver to promote innovation.

On February 11, the World Government Summit opened in Dubai, United Arab Emirates. Baidu founder Robin Li stated at the conference that in the field of AI or the IT industry, most innovations are related to cost reduction. If costs are reduced by a certain proportion, productivity also increases by the same proportion.

"Today, the speed of innovation is much faster than ever before. According to Moore's Law, every 18 months, performance doubles while prices halve. Nowadays, the inference cost of large models is reduced by more than 90% annually," said Robin Li.

02

A Low-Cost Storm Has Already Been Ignited

Compared to cloud vendors that have announced the listing of DeepSeek and their official list prices for invocation, Baidu Intelligent Cloud services offer significant advantages. Simultaneously, Baidu Intelligent Cloud is further enriching the AI model ecosystem on the platform, providing users with more diverse and powerful model options.

Overall, compared to other vendors, Baidu has proposed a highly "cost-effective" solution tailored to current market needs, helping users achieve a balance between product effectiveness, performance, and cost.

Baidu Intelligent Cloud Qianfan ModelBuilder is a platform related to large models launched by Baidu Intelligent Cloud, providing services such as model invocation and model effect tuning. It offers cost-effective Ernie Bot models and open-source model services, along with a one-stop tool chain for model effect tuning, encompassing data processing, model fine-tuning, model evaluation, and model quantization.

After DeepSeek broke the high-cost barrier dominated by capital and computing power in global AI large models with Chinese efficiency, the overall industry ecosystem has undergone significant changes.

Compared to vendors that have announced the listing of DeepSeek and the prices for direct invocation through official channels, Baidu Intelligent Cloud offers a 50% discount on R1 invocation compared to the official list price and a 30% discount on V3 invocation, the lowest across the entire network.

It can be said that, under the catalysis of DeepSeek, obtaining a reliable product experience at a reasonable cost is more important than blindly "burning money" and stacking computing power.

Under the current approach, Baidu Intelligent Cloud's products and services have already achieved remarkable results in the market.

According to official data, Baidu Intelligent Cloud's Qianfan large model platform has helped customers fine-tune 33,000 models and develop 770,000 enterprise applications. These applications cover various fields such as finance, government affairs, automobiles, and Internet technology, providing enterprise customers with a convenient, efficient, and economical large model usage and development experience.

From an industry perspective, Baidu's cost-effective solution not only balances its own investment costs but also provides new ideas and directions for the entire industry.

03

The Technological "Drag Race" of AI Large Models

Observing global AI competition, in addition to maintaining lower-cost computing power, achieving higher performance is also crucial for related enterprises to ensure their leading position.

Effectively reducing model invocation prices and providing more cost-effective solutions cannot be achieved without Baidu Intelligent Cloud's powerful and efficient computing power support, as well as its deep integration in reasoning engine performance optimization technology, reasoning service engineering architecture innovation, and full-link security guarantees for reasoning services.

Among these, Baidu's self-developed 10,000-card cluster is the key to achieving cost reduction in computing power.

Looking at the global large model competition landscape, it is evident that a single cluster with 10,000 cards has become an indispensable configuration for related enterprises.

After all, a 10,000-card cluster can continuously reduce the training cycle of models with 100 billion parameters, enabling rapid iteration of AI native applications. Simultaneously, the 10,000-card cluster supports multi-task concurrency capabilities. Through dynamic resource partitioning, a single cluster can simultaneously train multiple lightweight models, improving cluster utilization and achieving exponential reductions in training costs through communication optimization and fault tolerance mechanisms.

In early February, Baidu Intelligent Cloud announced the lighting up of the Kunlun 3rd-generation 10,000-card cluster, the first domestically developed 10,000-card cluster to be officially lit up. The Kunlun 3rd-generation 10,000-card cluster not only provides solid computing power support for Baidu but is also expected to drive the trend of cost reduction in models.

In terms of reasoning engine performance, based on its technical accumulation in large model reasoning performance optimization, Baidu Intelligent Cloud has conducted extreme performance optimization for the computation of the DeepSeek model's MLA structure. Through effective overlapping of different resource type operators such as computation, communication, and memory, as well as an efficient Prefill/Decode separated reasoning architecture, it achieves a significant increase in model throughput and a notable reduction in model reasoning costs under the condition that the core latency indicators TTFT/TPOT meet SLA.

At the reasoning service level, Baidu Intelligent Cloud has also conducted in-depth optimization and innovation, conducting a strict performance comparison of the push/pull mode for the reasoning architecture. Simultaneously, Baidu Intelligent Cloud's verified pull mode demonstrates superior performance in key indicators such as request processing success rate, response latency, and throughput.

To further enhance system stability and user experience, Baidu Intelligent Cloud ingeniously designed a retry mechanism for failed requests, significantly enhancing the system's fault tolerance and service SLA compliance rate.

Additionally, addressing the repeated Prompt prefixes in scenarios such as multi-turn dialogues and system settings, Baidu Intelligent Cloud has implemented mainstream KV-Cache reuse technology, supplemented by a global Cache-aware traffic scheduling strategy. This initiative effectively avoids repeated calculations of Token KV, thereby significantly reducing reasoning latency and improving reasoning throughput.

Moreover, regarding the security guarantees that users are concerned about, the platform integrates exclusive content security operators based on Baidu's long-term accumulation of large model security technology, achieving model security enhancement and enterprise-level high availability guarantees.

Simultaneously, based on the data security and model protection mechanisms throughout the entire lifecycle of large models, the models on the Qianfan platform all enjoy robust security guarantees; special optimizations in security ensure higher security for enterprise users of the DeepSeek-R1 & DeepSeek-V3 models during use.

In today's competitive landscape of large models, the technological "drag race" of AI large models is intensifying. Possessing high-performance and lower-cost computing power remains an important means for enterprises to achieve product popularization and maintain their leading position.

Under the strategy of providing cost-effective services, Baidu Intelligent Cloud's commercialization is accelerating, and the business model of Baidu's large models is expected to be further refined.

Some images are sourced from the internet. Please inform us for deletion if there is any infringement.